The problem is the estimate that changes when another year arrives
A fisheries stock assessment may report the spawning biomass for a particular year, then revise that estimate when data from the following year are added. That is not automatically an error. A new survey, a revised age composition, a new catch observation or an improved estimate of selectivity can change what a model infers about earlier years. A retrospective analysis deliberately asks how estimates made at different assessment end dates disagree about the same historical quantities. The essential question is not whether the model’s newest estimate differs from an older publication. It is whether the sequence of revisions has a systematic direction and size that could distort decisions about the stock. The phrase retrospective pattern refers to a structured sequence of revisions, not simply a single surprise after one new survey.
A useful report starts with a simple comparison: what was estimated, for which year, using what data cutoff and model configuration? “Biomass in 2022 was 80,000 tonnes” is incomplete if a later assessment says 64,000 tonnes. Were those two numbers forecasts for 2022, estimates calculated with different terminal years, or outputs of substantially different methods? A classic retrospective peel holds the model specification broadly fixed while removing one or more terminal years of data and refitting the remaining series. Comparing each peel with the corresponding years in the full model isolates a particular pattern of data-window sensitivity. A model revision between assessment editions is a related but different exercise, and should not be silently pooled with routine peels.
What a peel actually removes
Imagine that the full assessment uses annual observations from 2005 through 2025. Its terminal year is 2025. A one-year peel removes the observations for 2025 and refits the model through 2024. A two-year peel fits through 2023, and successive peels may end in 2022 and 2021. Each run re-estimates relevant parameters from its shortened dataset. A valid comparison uses the full model’s estimate for the peel’s terminal year as the reference. The 2024 peel is compared with the full model’s estimate for 2024, not with the full model’s 2025 estimate. Comparing unlike years can turn a real time trend in biomass into an artificial diagnosis of assessment bias.
The operation may sound like hiding information from the model, but it serves a constructive purpose. Each shortened fit imitates the information situation that would have existed at an earlier end date while keeping other modelling choices stable. If shortened runs repeatedly estimate a larger stock than the full run eventually assigns to those same terminal years, a one-direction pattern is visible. That is a warning about sensitivity to new observations. It does not by itself identify whether the full fit, the old peels or neither is closest to the true stock. Model outputs are estimates; retrospective agreement is a diagnostic of stability, not an independent measurement of reality.
Reading the spaghetti plot instead of admiring it
Retrospective plots often show a thick line for the full assessment and several shorter lines for successive peels. The curves may look like a tangle because each peel traces biomass estimates for many years, not only its terminal year. Locate the rightmost point of every peel, read its year, then compare its vertical value with the full curve at that same year. Trace whether the peel curves tend to end above, below or around the full line. Record whether the differences are concentrated near terminal years or extend far back in time. This simple reading prevents a misleading conclusion based merely on how many lines appear to cross.
A plot with nearly overlapping curves suggests that the fitted history is relatively insensitive to truncating the newest observations under the tested specification. A plot whose peeled curves consistently sit above the full series near each terminal date exhibits a positive retrospective pattern under one common convention. The reverse alignment produces a negative pattern. Colour legends alone can hide the direction if the reader confuses model run year with the biological year being estimated. When rewriting a complex graph as plain language, specify which curve supplied each numerator and denominator, the years compared, and whether a “positive” sign means earlier estimates were higher or lower than their subsequent revisions.
Defining Mohn’s rho without losing the denominator
A conventional version of Mohn’s rho averages the relative terminal-year differences across several retrospective peels. For each peel ending in year y, calculate (estimate from that peel for y minus estimate from the full run for y) divided by the full-run estimate for y. Sum the selected relative differences and divide by the number of peels. Written compactly, rho is an average of peel-minus-full divided by full, using matched historical years. The object may be biomass, spawning-stock biomass, recruitment or fishing mortality, but the label must say which. Different implementations or assessment traditions sometimes specify slightly different conventions, so a source report should always provide its actual formula.
This is a relative diagnostic, not an absolute tonnage. A ratio of 0.20 is dimensionless and corresponds to a 20 percent average excess of the selected peeled estimates over corresponding full estimates under that definition. Its magnitude depends on the denominator: using the peeled estimate instead would produce a different numerical result. Do not replace each denominator with the last-year full biomass; each peel should use its own matched year’s full estimate. Similarly, do not call the arithmetic average an annual growth rate. The quantity summarizes how a family of model fits changes its mind about historical stock attributes when recent years of data are included or excluded.
A completely checkable three-peel calculation
Consider an illustrative assessment with three matched terminal years. For 2022, the shortened model estimates 120,000 tonnes of spawning biomass and the full model estimates 100,000 tonnes, yielding a relative difference of (120,000 – 100,000)/100,000 = 0.20. For 2023, the peel estimates 88,000 tonnes and the full model 80,000 tonnes, giving 0.10. For 2024, the peel estimates 72,000 tonnes and the full model 60,000 tonnes, giving 0.20. The mean of 0.20, 0.10 and 0.20 is approximately 0.1667, or 16.7 percent. Those numbers are invented for teaching and must not be reported as a real stock assessment.
The arithmetic is easy; the interpretation requires discipline. Under the stated convention, the peels were on average higher than later full-run estimates for the same years. This is a positive retrospective pattern in spawning biomass. It does not establish that every historical biomass estimate was high by exactly 16.7 percent. It does not show that the current terminal-year biomass is definitely 16.7 percent too high, nor does it justify mechanically reducing catches by 16.7 percent. It says that, among the selected historical peels, estimates revised downward on average as later information arrived. Any operational adjustment needs independent evidence and assessment of the effect on decisions.
Why the sign can reverse for fishing mortality
Biomass and fishing mortality can respond in different directions to a new observation. If a model previously thought there were many fish and later reconstructs fewer fish having produced similar catches, its inferred fishing mortality may rise. A retrospective diagnostic for F can therefore have a different sign from the corresponding diagnostic for biomass. The same generic formula must be applied to the F estimates with the same-year comparison, but the practical meaning changes. Earlier peels overestimating biomass is not numerically equivalent to earlier peels underestimating F by the same percentage. The system contains nonlinear survival, selection and catch equations.
For example, a peel might estimate F = 0.25 per year for 2023 while the full run estimates F = 0.30 per year for that same biological year. Under the peel-minus-full divided by full convention, the relative difference is (0.25 – 0.30)/0.30 = -0.1667. That negative F component can coexist with a positive biomass rho. Repeating the calculation across peels gives a pattern for F, which should be named separately. A report saying merely “rho was positive” is ambiguous unless it identifies the parameter, exact convention and sampled peels. Technical clarity protects against a reversed risk message.
Retrospective pattern versus retrospective bias
“Bias” sounds like a verdict that an estimator is systematically wrong relative to truth. Retrospective patterns are observable differences among model fits; truth is usually not observed independently at the same scale. A model can show a persistent pattern while its latest estimates remain closer to reality than the older peels. A model can also show little retrospective pattern while all versions share a common structural error. The important distinction was emphasized in fisheries methodological discussions, including work examining why retrospective patterns should not be interpreted uncritically as known assessment bias. Diagnosis of a pattern is not causal identification.
A helpful analogy is a succession of weather forecasts updated each day. You can measure how the forecast for Saturday changed when another day’s observations arrived. That revision history may reveal a tendency, but it does not prove which forecast was correct unless Saturday’s conditions can be independently observed. Fish biomass is harder to observe directly because survey catchability, age structure and recruitment all have uncertainties. A retrospective statistic therefore flags a modelling property to investigate. It is not a replacement for survey validation, management strategy evaluation or tests of alternative assumptions about biology and observation.
Structural causes that deserve investigation
One cause of changing historical estimates is an inaccurate assumption about natural mortality. If natural mortality genuinely changes with time but the model forces it to remain constant, the addition of recent data may cause the model to redistribute errors through estimated recruitment, fishing mortality or biomass. Another cause is incomplete modelling of survey catchability, especially if gear efficiency or fish distribution changes over time. Changing age or size selectivity, unrepresentative sampling, misspecified recruitment variation and errors in ageing fish can also generate systematic revisions. Any of these can make successive peels drift even when the optimizer finds a technically converged solution.
The diagnostic task is to ask which mechanism predicts the observed pattern. If biomass peels change predominantly after a survey series began, examine its calibration and weighting. If revisions coincide with changes in ageing protocols, evaluate age-reading matrices. If strong cohorts are repeatedly revised, inspect recruitment variance and composition likelihoods. If catches are incomplete, examine discards and delayed reporting. Do not select a single cause from the direction of rho alone. Multiple mechanisms can produce the same sign, and a small average can conceal large positive and negative differences that cancel one another.
Sensitivity to how many peels you choose
A retrospective statistic based on three peels need not agree with one based on seven. The newly included historical years may feature different recruitment regimes, data coverage or survey design. When reporting rho, state the number of peels, terminal years, fitted model version and parameter. It is also useful to display the individual relative differences rather than only their mean. A mean of zero can arise from differences of +20, -20 and zero percent, which tells a different stability story from three differences near zero. The trajectory and spread are as informative as the average.
Consider an illustrative set of five relative differences: 0.30, 0.25, 0.20, -0.10 and -0.15. Their mean is 0.10, even though the latest two peels disagree in sign with the earlier three. A decision maker who sees only “rho = 0.10” misses a possible shift in the model’s behaviour. Analysts should explore the peels chronologically, evaluate plausible changes in fishery structure, and check whether one exceptional year dominates the statistic. Any choice to trim peels should be documented before assessing whether it improves the story. Selective omission of inconvenient years is not a legitimate repair.
Statistical intervals and model risks are different layers
A model may report a narrow confidence or credibility interval around terminal biomass while showing a strong retrospective pattern. These findings are not logically inconsistent: the interval may quantify uncertainty within the assumed model, whereas the retrospective analysis challenges sensitivity to data-window choices and possibly structural assumptions. A model-based interval need not include the full range of misspecification risks. Conversely, retrospective differences can reflect genuine information gains rather than an omitted uncertainty component. Readers should ask which uncertainty the reported band describes before using it to support a probability statement.
A useful assessment can display several perspectives: conditional model intervals, retrospective peel differences, sensitivity fits under alternative natural mortality or selectivity assumptions, and results of simulation tests if available. These diagnostics should be compared, not added as percentages as though they represented independent variances. Saying “20 percent rho plus 15 percent sampling error equals 35 percent uncertainty” is not justified. The sources of variability may be dependent and have different meanings. The goal is a coherent account of what is known about the stock, rather than an inflated single number that appears conservative but has no defensible statistical interpretation.
What a retrospective correction actually does
Some assessment workflows apply an adjustment motivated by rho to terminal estimates. Such procedures are model- and management-specific and should not be described as automatically correcting bias. If the unadjusted terminal biomass is 100,000 tonnes and a rule reduces it by a chosen proportion, the report must give the exact rule and specify whether the factor is rho, 1/(1 + rho), or something else. Different transformations of a 20 percent diagnostic do not produce identical values. A reduction to about 83,333 tonnes from dividing by 1.20 differs from an 80,000-tonne result obtained by multiplying by 0.80.
There are practical reasons to be cautious. A correction may improve short-term advice in one simulated situation and worsen it in another. It may introduce a discontinuity between the final historical estimate and the adjusted terminal figure, making the series difficult to read. It may conceal structural problems that remain unresolved when the next survey arrives. Peer-reviewed analyses have compared model ensemble approaches and terminal corrections, showing that neither simple method identifies the source of a pattern. Whenever an adjustment is used, document its evidence, risk consequences, and relationship to the harvest control rule rather than presenting it as a mathematical cure.
The difference between a diagnostic and management advice
Management advice depends on the probability of harmful outcomes, the stock’s biological reference points, current status, recruitment and future fishing pressure. A retrospective pattern matters because it may undermine confidence in biomass and F estimates used in projections. But a positive rho is not itself an annual catch limit, an overfishing threshold or a legal status determination. Those outputs require additional assessment and management procedures. When communicating to fishers or the public, distinguish the statement “historical estimates were repeatedly revised downward” from the statement “this stock is currently overfished.” The first may motivate further investigation; it does not logically entail the second.
A sensible response to a substantial pattern is to review possible model misspecification, examine independent indices and evaluate how alternative advice performs under uncertainty. Regulators may apply precautionary buffers according to their rules, but those buffers should not be reverse-engineered solely from the sign of one statistic. The key is to protect decision quality while acknowledging what the diagnostic cannot settle. The same principle helps readers understand why assessments may change after new information without concluding that earlier scientists acted improperly. Transparent revision is an expected part of evidence-based inference.
From formulas into clear sentences
A high-quality explanation should translate the algebra into a claim that a non-specialist can test. For example: “Across four annual retrospective peels, biomass estimates at the end of each shortened fit averaged 18 percent higher than the later full-fit estimates for those same years.” This gives the parameter, number of peels, sign, denominator and comparison. Add that it is a model-stability diagnostic rather than independently verified biomass bias. If the full model was rerun with changed data conventions, state that fact before interpreting the difference as a clean retrospective peel experiment.
Avoid saying “scientists discovered fish were 18 percent fewer” unless independent evidence supports that conclusion. The revision could come from model assumptions or the changing information set. Avoid saying “the model is wrong by 18 percent”; a retrospective mean does not establish truth. Avoid reporting a positive rho without specifying whether it belongs to biomass, recruitment or F. The strongest report provides the individual matched-year estimates and enough methodological detail that another reader can calculate each relative difference, inspect sign conventions and understand why management advice might or might not change.
Worked example: cancellation can conceal a serious pattern
Suppose three consecutive biomass peels have relative differences of +0.40, -0.30 and -0.10. Their arithmetic mean is zero. A report that publishes only rho = 0.00 might lead an uninformed reader to infer that assessments were very stable. In fact the differences were large in both directions. One peel was 40 percent higher than its matched full-fit estimate; another was 30 percent lower. Displaying the signed values, absolute values and relevant dates reveals information the mean suppresses. The mean absolute deviation here is about 0.267, not zero, although it measures a different property from the signed rho.
The example also warns against trying to choose between a large rho and a small rho as though the smallest number always proved the best model. A model can reduce mean rho by making opposite-signed changes that cancel, while continuing to fit some observations poorly. The diagnostic should be read alongside residuals, survey indices, composition fits, age-specific patterns and biological plausibility. If a team tunes its model only to minimize rho, it risks optimizing a symptom rather than explaining the data. The purpose is to improve the model’s reliability, not to win a contest for the tidiest retrospective graph.
Historical biomass is not identical to contemporary abundance
A spawning biomass estimate is an inference about the weight of mature fish under specified maturity schedules, weight-at-age measurements and abundance estimates. It is not simply the number of fish counted on a boat. When weights or maturity schedules change, reconstructed historical spawning biomass can shift even without the underlying numbers-at-age moving in the same proportion. That makes clear parameter labeling essential. “Biomass,” “spawning stock biomass,” “numbers at age” and “recruitment” identify different quantities with different units and biological meanings.
Consider a year with an estimated one million mature fish. If the mean weight represented in the assessment is two kilograms per fish, the corresponding illustrative mature biomass is 2,000 tonnes. If revised age structure and weights imply an average of 1.8 kilograms without a change in the number of mature fish, the estimate becomes 1,800 tonnes. This arithmetic is not itself a retrospective peel, but it illustrates why readers must inspect which component changed. If the question is whether mortality or recruitment changed, a biomass revision alone cannot answer it. A careful technical note identifies the estimated biological object and the input assumptions that define its value.
Retrospective differences can be localized in time
Some plots show that successive peels agree well for early years yet depart sharply near the terminal year. This can happen because data near the end of the time series have fewer future observations to constrain their interpretation. Age cohorts have not yet been seen at as many older ages, surveys have not yet revealed their later trajectories, and some catches or discards may be reported with delays. A larger terminal spread may therefore reflect weaker identification rather than a fully systematic error throughout the stock history. The location of disagreement helps focus investigation.
Other plots show revisions reaching decades into the past. That may suggest parameter confounding or changes in a parameter shared across the whole series, such as selectivity or growth. Reading only the last dot of each peel would miss this deeper instability. A practical presentation can provide a main retrospective chart plus a small table of terminal differences and a concise note about older-year effects. The two views answer different questions: whether recent advice is sensitive to new data and whether the estimated historical population trajectory also changes. Neither question should be compressed into rho alone.
Test the role of an independent abundance index
A fishery-independent survey can constrain population estimates, but its index is not automatically proportional to true biomass through time. Its relationship may involve catchability, survey footprint, season, gear efficiency and fish distribution. If survey operations change, the integrated assessment may revise historical stock sizes when new index observations arrive. A diagnostic experiment can refit variants with alternative index weighting, time blocks or justified catchability models, then compare retrospective patterns. The goal is to distinguish information from measurement artefact, not to discard a survey merely because its newest value disagrees with the model.
Suppose the survey index falls sharply while commercial catch remains stable. One interpretation is declining stock abundance with fishing pressure maintained. Another is a spatial distribution change moving fish outside the survey stations, while the fishery follows them. Still another is changed survey catchability. The retrospective peel cannot by itself select among these accounts. External evidence from habitat, tagging, survey calibration or neighbouring indices becomes valuable. A clear translation should say what each index measures and why its relationship to stock abundance is thought to remain sufficiently stable. This is the same denominator discipline needed when reading fisheries CPUE and relative abundance.
An evaluation must use plausible simulated worlds
A management strategy evaluation can test how candidate assessment and harvest rules perform across simulated stocks with known underlying dynamics. In that setting the analyst can compare estimated biomass with a simulated truth, something unavailable for a real stock. By introducing different recruitment variation, observation noise, selectivity changes or natural mortality patterns, the exercise asks whether a chosen retrospective adjustment improves or worsens the decisions a manager would make. Repeating simulations many times reveals the frequency of undesirable outcomes, not just one convenient example. That distinction matters because a method that seems helpful in a single historical plot may fail when the underlying cause is different.
The simulated worlds must themselves be scrutinized. A test covering only one type of model misspecification may exaggerate confidence that a correction is robust. If the true stock’s age structure, biology or observation process lies outside the explored range, performance claims may not transfer. Evaluation results should report scenarios, their plausibility, and measurable outcomes such as the probability of falling below a biomass limit, average catch, catch variability and rebuilding time. State whether a proposed correction was evaluated before advocating it. An adjustment derived from rho is a decision method, not a neutral fact about the data.
Alternative models are evidence only when compared fairly
When investigators try a time-varying M, a different selectivity curve or an alternative abundance-index weighting, a reduced retrospective pattern is encouraging but not sufficient. The alternative model may fit one component better while worsening others, create unidentifiable parameters or rely on unsupported biological assumptions. A fair comparison includes objective functions or likelihood diagnostics where appropriate, residual plots, predictive checks and stability under small data changes. It also considers whether a proposed mechanism would plausibly operate in the real fishery and whether independent observations support it. Model complexity can conceal uncertainty rather than resolve it.
For example, giving every year an independent catchability parameter could eliminate a survey fit conflict, but it may also absorb genuine biomass trends and leave the abundance scale poorly identified. Conversely, forcing catchability to remain constant despite a documented gear change can create systematic residuals and retrospective drift. A defensible middle approach might model a known survey transition explicitly, supported by calibration data. The best explanation is therefore causal and testable: specify what changed in the data-generating process and how that predicted change affects estimates. The direction of rho is a clue, not a full explanation.
Transfer exercise: calculate a four-peel rho
A fictional stock assessment offers the following peel and full-fit spawning biomass pairs, all in thousands of tonnes and matched by year: 2020 has 75 and 60; 2021 has 66 and 60; 2022 has 54 and 60; 2023 has 72 and 60. Compute the four relative differences using full fit in the denominator. They are 0.25, 0.10, -0.10 and 0.20. Their sum is 0.45, and dividing by four yields rho = 0.1125, or 11.25 percent. The individual values should appear in the answer, because reporting the mean alone hides the one negative peel.
Now explain what can and cannot be concluded. The selected shortened fits were higher on average than the later full fit for matching historical years, by the specified relative convention. One peel went in the opposite direction. Nothing in the calculation establishes true stock biomass, a legal stock status or a justified catch reduction. A defensible report would ask whether the model assumptions stayed fixed, whether those four years are representative and whether the size of the differences affects management decisions. Any answer that writes “11.25 percent more fish died” confuses retrospective model revision with mortality and fails the conceptual test.
Transfer exercise: reverse the denominator
Take a matched pair where a peel estimates biomass of 110,000 tonnes and the full run estimates 100,000 tonnes. The usual peel-minus-full divided by full calculation gives +0.10. If a trainee mistakenly divides the same 10,000-tonne difference by the peeled 110,000 tonnes, the result is about +0.0909. Both numbers describe a relative difference, but they answer different denominator questions. They are not interchangeable when combining many peels or comparing published statistics. A difference of 20 percent under one convention becomes approximately 16.7 percent under the other.
For a second check, suppose the peel estimates 90,000 tonnes while the full run estimates 100,000 tonnes. The conventional ratio is -0.10, and the alternative peel-denominator result is approximately -0.1111. The asymmetry is a useful warning that sign and percentage are not self-defining. In a translation from formula to prose, explicitly say “relative to the full-model estimate for the same year.” If original software or documentation uses another formula, report that convention faithfully rather than silently converting the result while keeping its old label.
Transfer exercise: time-window and model-change confusion
A full model through 2025 reports biomass for 2023 as 52,000 tonnes. Its two-year peel ending in 2023 reports 64,000 tonnes for 2023. Under the conventional matched-year formula the difference is (64,000 – 52,000)/52,000, about +0.2308. A separate older assessment published in 2023 reports 70,000 tonnes, but it used a different growth curve and a revised catch series. Can its 70,000-tonne figure be substituted for the two-year peel in the original rho calculation? No. It is not a clean truncation of the same model and data conventions.
The proper analysis labels two distinct comparisons. The 64,000-versus-52,000 contrast belongs to the fixed-specification retrospective exercise. The 70,000-versus-52,000 difference is an inter-assessment revision mixing new data with changed modelling assumptions. Both may matter for trust and communication, but combining them without distinction destroys the meaning of the statistic. A technical note should preserve model version, data cutoff, year of the biological estimate, source of the catch series and the specific parameter compared. Keeping these identifiers is often more important than adding another decimal place.
Frequently asked questions about retrospective results
Does rho measure the true percentage error in the stock assessment? No. It measures selected differences among refitted assessments under a stated formula. Does positive biomass rho prove overfishing? No. Fishing pressure and biomass status need separate biological and management evidence. Can a negative rho be good news? Not necessarily. Its sign describes the direction of revisions, not their accuracy or desirability. Does rho equal the growth rate of the stock? No. The numerator compares model estimates for the same year, not biomass in successive biological years. Is a small rho sufficient evidence that a model is reliable? No. Shared structural bias may persist even when successive peels agree.
What should a report include with rho? At minimum, the stock, parameter and units; full model end year; number and years of peels; individual matched-year values; formula and denominator; sign convention; model version; and interpretation limits. Why may independent scientists disagree about adjustment? Because adjustments interact with error mechanisms, forecast risks, harvest rules and competing fishery objectives. Is there one universal acceptable rho cutoff? There is no context-free threshold that converts the statistic into guaranteed model correctness. Evaluation should consider species, data quality, decision consequences and the diagnostic’s sampling behaviour.
How this connects to other fisheries figures
Retrospective analysis fits into a wider reasoning chain. Catch-at-age data and survival accounting influence fitted numbers of fish. Ageing and maturity schedules help determine spawning biomass. Surveys provide abundance signals with their own observation error. Stock–recruitment relationships describe possible offspring production and uncertainty. Projections convert the assessment state into risk under alternative future catches. A retrospective warning can influence confidence in those projections, but it is not a substitute for them. For related techniques, see eduKate’s catch-at-age VPA explanation, stock–recruitment guide and surplus-production guide.
An evidence chain should also keep roles clear. The model estimates latent stock states, retrospective peels diagnose sensitivity, independent observations challenge mechanisms, and fishery authorities choose management measures under applicable rules. Calling all these steps “the assessment result” hides where judgement enters. A careful reader can ask for the data that would alter a conclusion, not merely the name of the algorithm. That makes scientific explanation useful to managers, students and the public without pretending that uncertainty has been eliminated.
Final reading rule: revisions must be legible
When an assessment revises biomass downward, the communicative obligation is to show what changed and why the new estimate is considered preferable or at least better informed. When revisions happen repeatedly in one direction, display the pattern instead of burying it in a dense appendix. When a statistic summarizes it, disclose the formula and the individual differences. When an adjustment is proposed, test the decision consequences rather than treating the arithmetic as proof. The reader should be able to understand whether an uncertainty comes from a measurement, a biological process, an assessment assumption or a management choice.
This approach is useful far beyond fisheries. Whenever an evolving model re-estimates the past as new observations arrive, distinguish revisions to estimates from changes in the underlying historical events. A model changing its mind is information about the evidence system. The event itself did not change retroactively. Scientific translation succeeds when it makes that difference explicit, keeps denominators and time labels intact, and gives readers a reproducible calculation together with an honest explanation of what remains unknown.
Primary sources and methodological reading
NOAA repository: Legault and colleagues on Rose versus Rho approaches. NOAA repository: statistical behaviour of retrospective patterns. Cadrin, 2025, Misinterpreting retrospective patterns in fishery stock assessment. Deroba, 2014, consequences of applying Mohn’s rho. These works inform the cautions about interpretation and adjustment; all numerical exercises in this article are original hypothetical teaching examples.